@falai/agent 1.1.3 → 1.2.1

package/src/core/Agent.ts

@@ -18,6 +18,7 @@ import type {
   ValidationError,
   ValidationResult,
   AiProvider,
+  CompactionOptions,
 } from "../types";
 import { CompositionMode } from "../types";
 import type { StreamOptions, GenerateOptions, RespondParams } from "./ResponseModal";
@@ -35,9 +36,11 @@ import { Step } from "./Step";
 import { PersistenceManager } from "./PersistenceManager";
 import { SessionManager } from "./SessionManager";
 import { RoutingEngine } from "./RoutingEngine";
+import { PromptSectionCache } from "./PromptSectionCache";
 
 import { ResponseModal } from "./ResponseModal";
 import { ToolManager } from "./ToolManager";
+import { CompactionEngine } from "./CompactionEngine";
 
 /**
  * Error thrown when data validation fails
@@ -78,6 +81,8 @@ export class Agent<TContext = any, TData = any> {
   private _knowledgeBase: Record<string, unknown> = {};
   private _schema?: StructuredSchema;
   private _collectedData: Partial<TData> = {};
+  private _compactionOptions?: CompactionOptions;
+  private _promptSectionCache: PromptSectionCache;
 
   /** Public session manager for easy session management */
   public session: SessionManager<TData>;
@@ -125,9 +130,14 @@ export class Agent<TContext = any, TData = any> {
     // Initialize current session if provided
     this._currentSession = options.session;
 
+    // Initialize prompt section cache
+    this._promptSectionCache = new PromptSectionCache(options.promptCache);
+
     // Initialize routing engine
     this._routingEngine = new RoutingEngine<TContext, TData>({
       routeSwitchMargin: options.routeSwitchMargin,
+      onRouteSwitch: () => this.invalidateRouteSections(),
+      promptSectionCache: this._promptSectionCache,
     });
 
     // Initialize ResponseModal for handling all response generation
@@ -205,6 +215,20 @@ export class Agent<TContext = any, TData = any> {
       this._knowledgeBase = { ...options.knowledgeBase };
     }
 
+    // Initialize compaction options if configured
+    if (options.compaction && options.compaction.enabled !== false) {
+      const compactionOptions: CompactionOptions = {
+        maxTokens: options.compaction.maxTokens,
+        compactionThreshold: options.compaction.compactionThreshold ?? 0.8,
+        preserveRecentCount: options.compaction.preserveRecentCount ?? 4,
+        maxToolResultChars: options.compaction.maxToolResultChars ?? 5000,
+        provider: options.provider,
+      };
+      CompactionEngine.validateOptions(compactionOptions);
+      this._compactionOptions = compactionOptions;
+      logger.debug("[Agent] Compaction options initialized and validated");
+    }
+
     // Initialize session manager with reference to this agent for bidirectional sync
     this.session = new SessionManager<TData>(this._persistenceManager, this);
 
@@ -505,6 +529,13 @@ export class Agent<TContext = any, TData = any> {
     this.options.routeSwitchMargin = value;
   }
 
+  /**
+   * Get the prompt section cache instance
+   */
+  get promptSectionCache(): PromptSectionCache {
+    return this._promptSectionCache;
+  }
+
   /**
    * Get the maximum steps per batch
    * @default 1
@@ -620,6 +651,7 @@ export class Agent<TContext = any, TData = any> {
    */
   set currentSession(value: SessionState | undefined) {
     this._currentSession = value;
+    this._promptSectionCache.invalidateAll();
   }
 
   // ---------------------------------------------------------------------------
@@ -705,6 +737,7 @@ export class Agent<TContext = any, TData = any> {
    */
   setCurrentSession(session: SessionState): void {
     this.currentSession = session;
+    this._promptSectionCache.invalidateAll();
   }
 
   /**
@@ -713,6 +746,18 @@ export class Agent<TContext = any, TData = any> {
    */
   clearCurrentSession(): void {
     this._currentSession = undefined;
+    this._promptSectionCache.invalidateAll();
+  }
+
+  /**
+   * Invalidate route-dependent prompt cache sections.
+   * Called automatically when the active route changes.
+   */
+  invalidateRouteSections(): void {
+    this._promptSectionCache.invalidate('activeRoutes');
+    this._promptSectionCache.invalidate('routeRules');
+    this._promptSectionCache.invalidate('routeProhibitions');
+    this._promptSectionCache.invalidate('routeKnowledgeBase');
   }
 
   /**
@@ -729,6 +774,13 @@ export class Agent<TContext = any, TData = any> {
     return this._persistenceManager !== undefined;
   }
 
+  /**
+   * Get the resolved compaction options (if compaction is configured)
+   */
+  getCompactionOptions(): CompactionOptions | undefined {
+    return this._compactionOptions;
+  }
+
   // ---------------------------------------------------------------------------
   // Core methods
   // ---------------------------------------------------------------------------
@@ -876,6 +928,10 @@ export class Agent<TContext = any, TData = any> {
     if (this.options.hooks?.onContextUpdate && previousContext !== undefined) {
       await this.options.hooks.onContextUpdate(this._context, previousContext);
     }
+
+    // Invalidate context-dependent prompt cache sections
+    this._promptSectionCache.invalidate('agentMeta');
+    this._promptSectionCache.invalidate('knowledgeBase');
   }
 
   /**
@@ -939,7 +995,7 @@ export class Agent<TContext = any, TData = any> {
   /**
    * Generate a response based on history and context as a stream
    */
-  async *respondStream(params: RespondParams<TContext, TData>): AsyncGenerator<AgentResponseStreamChunk<TData>> {
+  async * respondStream(params: RespondParams<TContext, TData>): AsyncGenerator<AgentResponseStreamChunk<TData>> {
     // Delegate to ResponseModal
     yield* this._responseModal.respondStream(params);
   }
@@ -1210,7 +1266,7 @@ export class Agent<TContext = any, TData = any> {
    * Modern streaming API - simple interface like chat() but returns a stream
    * Automatically manages conversation history through the session
    */
-  async *stream(
+  async * stream(
     message?: string,
     options?: StreamOptions<TContext>
   ): AsyncGenerator<AgentResponseStreamChunk<TData>> {

package/src/core/BatchPromptBuilder.ts

@@ -14,6 +14,7 @@ import type { Event } from '../types/history';
 import type { Route } from './Route';
 import { render, renderMany, createTemplateContext } from '../utils/template';
 import { PromptComposer } from './PromptComposer';
+import type { PromptSectionCache } from './PromptSectionCache';
 
 /**
  * Parameters for building a batch prompt
@@ -56,6 +57,8 @@ export interface BatchPromptResult {
  * **Validates: Requirements 4.1, 4.2, 4.3, 4.4**
  */
 export class BatchPromptBuilder<TContext = unknown, TData = unknown> {
+  constructor(private readonly promptSectionCache?: PromptSectionCache) { }
+
   /**
    * Build a combined prompt for a batch of Steps
    * 
@@ -89,7 +92,7 @@ export class BatchPromptBuilder<TContext = unknown, TData = unknown> {
     }
 
     // Build the combined prompt using PromptComposer for consistency
-    const composer = new PromptComposer<TContext, TData>(templateContext);
+    const composer = new PromptComposer<TContext, TData>(templateContext, this.promptSectionCache);
 
     // Add agent meta information
     await composer.addAgentMeta(agentOptions);

package/src/core/CompactionEngine.ts

@@ -0,0 +1,318 @@
+/**
+ * CompactionEngine - Manages conversation history size through multi-layered compaction strategies.
+ *
+ * Strategies are applied in order of cost:
+ * 1. none - history is under threshold, no action needed
+ * 2. tool_result_budget - truncate oversized tool results (no LLM call)
+ * 3. micro_compact - compress verbose tool outputs inline (no LLM call)
+ * 4. auto_compact - summarize old messages via LLM provider
+ */
+
+import log from "loglevel";
+import type { HistoryItem } from "../types/history";
+import type { CompactionOptions, CompactionResult } from "../types/compaction";
+
+export class CompactionEngine {
+    /**
+     * Validate CompactionOptions. Throws on invalid values.
+     */
+    static validateOptions(options: CompactionOptions): void {
+        if (
+            typeof options.compactionThreshold !== "number" ||
+            options.compactionThreshold < 0.5 ||
+            options.compactionThreshold > 0.95
+        ) {
+            throw new Error(
+                `compactionThreshold must be between 0.5 and 0.95, got ${options.compactionThreshold}`
+            );
+        }
+        if (
+            typeof options.preserveRecentCount !== "number" ||
+            options.preserveRecentCount < 2
+        ) {
+            throw new Error(
+                `preserveRecentCount must be >= 2, got ${options.preserveRecentCount}`
+            );
+        }
+        if (
+            typeof options.maxToolResultChars !== "number" ||
+            options.maxToolResultChars <= 0
+        ) {
+            throw new Error(
+                `maxToolResultChars must be > 0, got ${options.maxToolResultChars}`
+            );
+        }
+    }
+
+    /**
+     * Estimate token count using a character-based heuristic (~4 chars/token).
+     *
+     * For each HistoryItem: count content length + name length (if present) + 4 (role overhead).
+     * Returns Math.ceil(totalChars / 4). Empty history returns 0.
+     * Deterministic for the same input.
+     */
+    static estimateTokens(history: HistoryItem[]): number {
+        if (history.length === 0) return 0;
+
+        let totalChars = 0;
+        for (const item of history) {
+            const content = item.content;
+            const contentLength =
+                typeof content === "string" ? content.length : JSON.stringify(content ?? "").length;
+            totalChars += contentLength;
+            if ("name" in item && typeof item.name === "string") {
+                totalChars += item.name.length;
+            }
+            totalChars += 4; // role overhead
+        }
+        return Math.ceil(totalChars / 4);
+    }
+
+    /**
+     * Truncate tool results that exceed the per-message character budget.
+     *
+     * For HistoryItem with role === 'tool' whose content exceeds maxChars,
+     * truncate to maxChars and append truncation notice.
+     * Items within budget are returned unchanged.
+     */
+    static applyToolResultBudget(
+        history: HistoryItem[],
+        maxCharsPerResult: number
+    ): HistoryItem[] {
+        return history.map((item) => {
+            if (item.role !== "tool") return item;
+
+            const contentStr =
+                typeof item.content === "string"
+                    ? item.content
+                    : JSON.stringify(item.content ?? "");
+            const totalChars = contentStr.length;
+
+            if (totalChars <= maxCharsPerResult) return item;
+
+            const preview = contentStr.slice(0, maxCharsPerResult);
+            const truncatedContent = `${preview}\n\n[Truncated: ${totalChars} chars total, showing first ${maxCharsPerResult}]`;
+            return { ...item, content: truncatedContent };
+        });
+    }
+
+    /**
+     * Micro-compact: compress verbose tool outputs inline.
+     * Strips excessive whitespace and shortens JSON-like content in tool results.
+     * Preserves the last `preserveCount` messages unchanged.
+     */
+    private static microCompact(
+        history: HistoryItem[],
+        preserveCount: number
+    ): HistoryItem[] {
+        const cutoff = Math.max(0, history.length - preserveCount);
+        const compactable = history.slice(0, cutoff);
+        const preserved = history.slice(cutoff);
+
+        const compacted = compactable.map((item) => {
+            if (item.role !== "tool") return item;
+
+            const contentStr =
+                typeof item.content === "string"
+                    ? item.content
+                    : JSON.stringify(item.content ?? "");
+
+            // Strip excessive whitespace: collapse runs of whitespace to single space
+            const compressed = contentStr
+                .replace(/\s+/g, " ")
+                .trim();
+
+            return { ...item, content: compressed };
+        });
+
+        return [...compacted, ...preserved];
+    }
+
+    /**
+     * Count how many messages differ between original and compacted history.
+     */
+    private static countDifferences(
+        original: HistoryItem[],
+        compacted: HistoryItem[]
+    ): number {
+        let count = 0;
+        const len = Math.min(original.length, compacted.length);
+        for (let i = 0; i < len; i++) {
+            if (original[i] !== compacted[i]) count++;
+        }
+        // If lengths differ, count the extra messages
+        count += Math.abs(original.length - compacted.length);
+        return count;
+    }
+
+    /**
+     * Summarize old messages via LLM provider.
+     * On failure, returns null (caller should fall back).
+     */
+    private static async summarizeMessages(
+        messages: HistoryItem[],
+        options: CompactionOptions
+    ): Promise<string | null> {
+        try {
+            const messagesText = messages
+                .map((m) => {
+                    const content =
+                        typeof m.content === "string"
+                            ? m.content
+                            : JSON.stringify(m.content ?? "");
+                    return `[${m.role}]: ${content}`;
+                })
+                .join("\n");
+
+            const result = await options.provider.generateMessage({
+                prompt: `Summarize the following conversation concisely, preserving key facts, decisions, and context:\n\n${messagesText}`,
+                history: [],
+                context: {},
+                parameters: {
+                    maxOutputTokens: 1024,
+                    jsonSchema: {},
+                },
+            });
+
+            return result.message;
+        } catch {
+            return null;
+        }
+    }
+
+    /**
+     * Aggressive truncation fallback: remove oldest messages (no LLM needed).
+     * Keeps only the most recent messages that fit within the token budget.
+     */
+    private static aggressiveTruncate(
+        history: HistoryItem[],
+        options: CompactionOptions
+    ): HistoryItem[] {
+        const threshold = options.maxTokens * options.compactionThreshold;
+        const preserveCount = options.preserveRecentCount;
+
+        // Always preserve the last preserveRecentCount messages
+        const preserved = history.slice(-preserveCount);
+
+        // Try to keep as many older messages as fit within budget
+        const older = history.slice(0, -preserveCount);
+        const result: HistoryItem[] = [];
+
+        // Add older messages from most recent backwards until we'd exceed budget
+        for (let i = older.length - 1; i >= 0; i--) {
+            const candidate = [older[i], ...result, ...preserved];
+            if (CompactionEngine.estimateTokens(candidate) < threshold) {
+                result.unshift(older[i]);
+            } else {
+                break;
+            }
+        }
+
+        return [...result, ...preserved];
+    }
+
+    /**
+     * Multi-layered compaction strategy.
+     *
+     * Layer 1 (none): If estimatedTokens < maxTokens * compactionThreshold, return unchanged
+     * Layer 2 (tool_result_budget): Apply applyToolResultBudget, check if under threshold
+     * Layer 3 (micro_compact): Compress verbose tool outputs inline
+     * Layer 4 (auto_compact): Summarize old messages via LLM provider
+     *
+     * The last `preserveRecentCount` messages are NEVER modified or removed.
+     */
+    static async checkAndCompact(
+        history: HistoryItem[],
+        options: CompactionOptions
+    ): Promise<CompactionResult> {
+        CompactionEngine.validateOptions(options);
+
+        const threshold = options.maxTokens * options.compactionThreshold;
+        const estimatedTokens = CompactionEngine.estimateTokens(history);
+
+        // Layer 1: No compaction needed
+        if (estimatedTokens < threshold) {
+            return {
+                history,
+                strategy: "none",
+                estimatedTokens,
+                messagesCompacted: 0,
+            };
+        }
+
+        // Layer 2: Tool result budgeting (cheapest — no LLM call)
+        // Apply budget but preserve recent messages
+        const preserveCount = Math.min(options.preserveRecentCount, history.length);
+        const cutoff2 = Math.max(0, history.length - preserveCount);
+        const budgeted = [
+            ...CompactionEngine.applyToolResultBudget(
+                history.slice(0, cutoff2),
+                options.maxToolResultChars
+            ),
+            ...history.slice(cutoff2),
+        ];
+        let newEstimate = CompactionEngine.estimateTokens(budgeted);
+        if (newEstimate < threshold) {
+            return {
+                history: budgeted,
+                strategy: "tool_result_budget",
+                estimatedTokens: newEstimate,
+                messagesCompacted: CompactionEngine.countDifferences(history, budgeted),
+            };
+        }
+
+        // Layer 3: Micro-compaction (compress verbose tool outputs)
+        const microCompacted = CompactionEngine.microCompact(budgeted, preserveCount);
+        newEstimate = CompactionEngine.estimateTokens(microCompacted);
+        if (newEstimate < threshold) {
+            return {
+                history: microCompacted,
+                strategy: "micro_compact",
+                estimatedTokens: newEstimate,
+                messagesCompacted: CompactionEngine.countDifferences(
+                    history,
+                    microCompacted
+                ),
+            };
+        }
+
+        // Layer 4: Auto-compact (summarize old messages via LLM)
+        const oldMessages = microCompacted.slice(0, -preserveCount);
+        const recentMessages = microCompacted.slice(-preserveCount);
+
+        const summary = await CompactionEngine.summarizeMessages(
+            oldMessages,
+            options
+        );
+
+        if (summary !== null) {
+            const summaryItem: HistoryItem = {
+                role: "system",
+                content: `[Conversation Summary]\n${summary}`,
+            };
+            const finalHistory = [summaryItem, ...recentMessages];
+            return {
+                history: finalHistory,
+                strategy: "auto_compact",
+                estimatedTokens: CompactionEngine.estimateTokens(finalHistory),
+                messagesCompacted: oldMessages.length,
+                summary,
+            };
+        }
+
+        // Fallback: LLM summarization failed — aggressive truncation
+        log.warn(
+            "CompactionEngine: LLM summarization failed, falling back to aggressive truncation"
+        );
+        const truncated = CompactionEngine.aggressiveTruncate(
+            microCompacted,
+            options
+        );
+        return {
+            history: truncated,
+            strategy: "auto_compact",
+            estimatedTokens: CompactionEngine.estimateTokens(truncated),
+            messagesCompacted: history.length - truncated.length,
+        };
+    }
+}